OU/CWE-ou-minkata
1
0
Fork 0
mirror of https://foundry.openuru.org/gitblit/r/CWE-ou-minkata.git synced 2025-07-19 11:49:09 +00:00
Code Releases Activity

Update hsMatrix44 to use hsCpuID dispatcher.

Browse Source
This commit is contained in:
Joseph Davies
2012-04-22 14:51:39 -07:00
parent 42f4eec1f7
commit 304d15acfe
2 changed files with 61 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
Sources/Plasma/CoreLib/hsMatrix44.cpp
View File

@ -47,13 +47,16 @@ You can contact Cyan Worlds, Inc. by email legal@cyan.com
#include "hsStream.h" #include "hsStream.h"
#include <math.h> #include <math.h>
#ifdef HAVE_SSE #ifdef HS_SIMD_INCLUDE
# include <smmintrin.h> # include HS_SIMD_INCLUDE
#endif #endif
static hsMatrix44 myIdent = hsMatrix44().Reset(); static hsMatrix44 myIdent = hsMatrix44().Reset();
const hsMatrix44& hsMatrix44::IdentityMatrix() { return myIdent; } const hsMatrix44& hsMatrix44::IdentityMatrix() { return myIdent; }
// CPU-optimized functions requiring dispatch
hsFunctionDispatcher<hsMatrix44::mat_mult_ptr> hsMatrix44::mat_mult(hsMatrix44::mat_mult_fpu, 0, 0, hsMatrix44::mat_mult_sse3);
/* /*
For the rotation: For the rotation:
<20> 2 2 <20> <20> 2 2 <20>
@ -96,9 +99,47 @@ void hsMatrix44::DecompRigid(hsScalarTriple &translate, hsQuat &rotate) const
rotate.QuatFromMatrix44(*this); rotate.QuatFromMatrix44(*this);
} }
#ifdef HAVE_SSE hsMatrix44 hsMatrix44::mat_mult_fpu(const hsMatrix44 &a, const hsMatrix44 &b)
{
hsMatrix44 c;
if( a.fFlags & b.fFlags & hsMatrix44::kIsIdent )
{
c.Reset();
return c;
}
if( a.fFlags & hsMatrix44::kIsIdent )
return b;
if( b.fFlags & hsMatrix44::kIsIdent )
return a;
c.fMap[0][0] = (a.fMap[0][0] * b.fMap[0][0]) + (a.fMap[0][1] * b.fMap[1][0]) + (a.fMap[0][2] * b.fMap[2][0]) + (a.fMap[0][3] * b.fMap[3][0]);
c.fMap[0][1] = (a.fMap[0][0] * b.fMap[0][1]) + (a.fMap[0][1] * b.fMap[1][1]) + (a.fMap[0][2] * b.fMap[2][1]) + (a.fMap[0][3] * b.fMap[3][1]);
c.fMap[0][2] = (a.fMap[0][0] * b.fMap[0][2]) + (a.fMap[0][1] * b.fMap[1][2]) + (a.fMap[0][2] * b.fMap[2][2]) + (a.fMap[0][3] * b.fMap[3][2]);
c.fMap[0][3] = (a.fMap[0][0] * b.fMap[0][3]) + (a.fMap[0][1] * b.fMap[1][3]) + (a.fMap[0][2] * b.fMap[2][3]) + (a.fMap[0][3] * b.fMap[3][3]);
c.fMap[1][0] = (a.fMap[1][0] * b.fMap[0][0]) + (a.fMap[1][1] * b.fMap[1][0]) + (a.fMap[1][2] * b.fMap[2][0]) + (a.fMap[1][3] * b.fMap[3][0]);
c.fMap[1][1] = (a.fMap[1][0] * b.fMap[0][1]) + (a.fMap[1][1] * b.fMap[1][1]) + (a.fMap[1][2] * b.fMap[2][1]) + (a.fMap[1][3] * b.fMap[3][1]);
c.fMap[1][2] = (a.fMap[1][0] * b.fMap[0][2]) + (a.fMap[1][1] * b.fMap[1][2]) + (a.fMap[1][2] * b.fMap[2][2]) + (a.fMap[1][3] * b.fMap[3][2]);
c.fMap[1][3] = (a.fMap[1][0] * b.fMap[0][3]) + (a.fMap[1][1] * b.fMap[1][3]) + (a.fMap[1][2] * b.fMap[2][3]) + (a.fMap[1][3] * b.fMap[3][3]);
c.fMap[2][0] = (a.fMap[2][0] * b.fMap[0][0]) + (a.fMap[2][1] * b.fMap[1][0]) + (a.fMap[2][2] * b.fMap[2][0]) + (a.fMap[2][3] * b.fMap[3][0]);
c.fMap[2][1] = (a.fMap[2][0] * b.fMap[0][1]) + (a.fMap[2][1] * b.fMap[1][1]) + (a.fMap[2][2] * b.fMap[2][1]) + (a.fMap[2][3] * b.fMap[3][1]);
c.fMap[2][2] = (a.fMap[2][0] * b.fMap[0][2]) + (a.fMap[2][1] * b.fMap[1][2]) + (a.fMap[2][2] * b.fMap[2][2]) + (a.fMap[2][3] * b.fMap[3][2]);
c.fMap[2][3] = (a.fMap[2][0] * b.fMap[0][3]) + (a.fMap[2][1] * b.fMap[1][3]) + (a.fMap[2][2] * b.fMap[2][3]) + (a.fMap[2][3] * b.fMap[3][3]);
c.fMap[3][0] = (a.fMap[3][0] * b.fMap[0][0]) + (a.fMap[3][1] * b.fMap[1][0]) + (a.fMap[3][2] * b.fMap[2][0]) + (a.fMap[3][3] * b.fMap[3][0]);
c.fMap[3][1] = (a.fMap[3][0] * b.fMap[0][1]) + (a.fMap[3][1] * b.fMap[1][1]) + (a.fMap[3][2] * b.fMap[2][1]) + (a.fMap[3][3] * b.fMap[3][1]);
c.fMap[3][2] = (a.fMap[3][0] * b.fMap[0][2]) + (a.fMap[3][1] * b.fMap[1][2]) + (a.fMap[3][2] * b.fMap[2][2]) + (a.fMap[3][3] * b.fMap[3][2]);
c.fMap[3][3] = (a.fMap[3][0] * b.fMap[0][3]) + (a.fMap[3][1] * b.fMap[1][3]) + (a.fMap[3][2] * b.fMap[2][3]) + (a.fMap[3][3] * b.fMap[3][3]);
return c;
}
#ifdef HS_SSE3
# define MULTBEGIN(i) \ # define MULTBEGIN(i) \
xmm[0] = _mm_loadu_ps(fMap[i]); xmm[0] = _mm_loadu_ps(a.fMap[i]);
# define MULTCELL(i, j) \ # define MULTCELL(i, j) \
xmm[1] = _mm_set_ps(b.fMap[3][j], b.fMap[2][j], b.fMap[1][j], b.fMap[0][j]); \ xmm[1] = _mm_set_ps(b.fMap[3][j], b.fMap[2][j], b.fMap[1][j], b.fMap[0][j]); \
xmm[j+2] = _mm_mul_ps(xmm[0], xmm[1]); xmm[j+2] = _mm_mul_ps(xmm[0], xmm[1]);
@ -107,24 +148,23 @@ void hsMatrix44::DecompRigid(hsScalarTriple &translate, hsQuat &rotate) const
xmm[7] = _mm_hadd_ps(xmm[4], xmm[5]); \ xmm[7] = _mm_hadd_ps(xmm[4], xmm[5]); \
xmm[1] = _mm_hadd_ps(xmm[6], xmm[7]); \ xmm[1] = _mm_hadd_ps(xmm[6], xmm[7]); \
_mm_storeu_ps(c.fMap[i], xmm[1]); _mm_storeu_ps(c.fMap[i], xmm[1]);
#endif #endif // HS_SSE3
hsMatrix44 hsMatrix44::operator*(const hsMatrix44& b) const hsMatrix44 hsMatrix44::mat_mult_sse3(const hsMatrix44 &a, const hsMatrix44 &b)
{ {
hsMatrix44 c; hsMatrix44 c;
#ifdef HS_SSE3
if( fFlags & b.fFlags & hsMatrix44::kIsIdent ) if( a.fFlags & b.fFlags & hsMatrix44::kIsIdent )
{ {
c.Reset(); c.Reset();
return c; return c;
} }
if( fFlags & hsMatrix44::kIsIdent ) if( a.fFlags & hsMatrix44::kIsIdent )
return b; return b;
if( b.fFlags & hsMatrix44::kIsIdent ) if( b.fFlags & hsMatrix44::kIsIdent )
return *this; return a;
#ifdef HAVE_SSE
__m128 xmm[8]; __m128 xmm[8];
MULTBEGIN(0); MULTBEGIN(0);
@ -154,28 +194,7 @@ hsMatrix44 hsMatrix44::operator*(const hsMatrix44& b) const
MULTCELL(3, 2); MULTCELL(3, 2);
MULTCELL(3, 3); MULTCELL(3, 3);
MULTFINISH(3); MULTFINISH(3);
#else #endif // HS_SSE3
c.fMap[0][0] = (fMap[0][0] * b.fMap[0][0]) + (fMap[0][1] * b.fMap[1][0]) + (fMap[0][2] * b.fMap[2][0]) + (fMap[0][3] * b.fMap[3][0]);
c.fMap[0][1] = (fMap[0][0] * b.fMap[0][1]) + (fMap[0][1] * b.fMap[1][1]) + (fMap[0][2] * b.fMap[2][1]) + (fMap[0][3] * b.fMap[3][1]);
c.fMap[0][2] = (fMap[0][0] * b.fMap[0][2]) + (fMap[0][1] * b.fMap[1][2]) + (fMap[0][2] * b.fMap[2][2]) + (fMap[0][3] * b.fMap[3][2]);
c.fMap[0][3] = (fMap[0][0] * b.fMap[0][3]) + (fMap[0][1] * b.fMap[1][3]) + (fMap[0][2] * b.fMap[2][3]) + (fMap[0][3] * b.fMap[3][3]);
c.fMap[1][0] = (fMap[1][0] * b.fMap[0][0]) + (fMap[1][1] * b.fMap[1][0]) + (fMap[1][2] * b.fMap[2][0]) + (fMap[1][3] * b.fMap[3][0]);
c.fMap[1][1] = (fMap[1][0] * b.fMap[0][1]) + (fMap[1][1] * b.fMap[1][1]) + (fMap[1][2] * b.fMap[2][1]) + (fMap[1][3] * b.fMap[3][1]);
c.fMap[1][2] = (fMap[1][0] * b.fMap[0][2]) + (fMap[1][1] * b.fMap[1][2]) + (fMap[1][2] * b.fMap[2][2]) + (fMap[1][3] * b.fMap[3][2]);
c.fMap[1][3] = (fMap[1][0] * b.fMap[0][3]) + (fMap[1][1] * b.fMap[1][3]) + (fMap[1][2] * b.fMap[2][3]) + (fMap[1][3] * b.fMap[3][3]);
c.fMap[2][0] = (fMap[2][0] * b.fMap[0][0]) + (fMap[2][1] * b.fMap[1][0]) + (fMap[2][2] * b.fMap[2][0]) + (fMap[2][3] * b.fMap[3][0]);
c.fMap[2][1] = (fMap[2][0] * b.fMap[0][1]) + (fMap[2][1] * b.fMap[1][1]) + (fMap[2][2] * b.fMap[2][1]) + (fMap[2][3] * b.fMap[3][1]);
c.fMap[2][2] = (fMap[2][0] * b.fMap[0][2]) + (fMap[2][1] * b.fMap[1][2]) + (fMap[2][2] * b.fMap[2][2]) + (fMap[2][3] * b.fMap[3][2]);
c.fMap[2][3] = (fMap[2][0] * b.fMap[0][3]) + (fMap[2][1] * b.fMap[1][3]) + (fMap[2][2] * b.fMap[2][3]) + (fMap[2][3] * b.fMap[3][3]);
c.fMap[3][0] = (fMap[3][0] * b.fMap[0][0]) + (fMap[3][1] * b.fMap[1][0]) + (fMap[3][2] * b.fMap[2][0]) + (fMap[3][3] * b.fMap[3][0]);
c.fMap[3][1] = (fMap[3][0] * b.fMap[0][1]) + (fMap[3][1] * b.fMap[1][1]) + (fMap[3][2] * b.fMap[2][1]) + (fMap[3][3] * b.fMap[3][1]);
c.fMap[3][2] = (fMap[3][0] * b.fMap[0][2]) + (fMap[3][1] * b.fMap[1][2]) + (fMap[3][2] * b.fMap[2][2]) + (fMap[3][3] * b.fMap[3][2]);
c.fMap[3][3] = (fMap[3][0] * b.fMap[0][3]) + (fMap[3][1] * b.fMap[1][3]) + (fMap[3][2] * b.fMap[2][3]) + (fMap[3][3] * b.fMap[3][3]);
#endif
return c; return c;
} }

11
Sources/Plasma/CoreLib/hsMatrix44.h
View File

@ -44,6 +44,7 @@ You can contact Cyan Worlds, Inc. by email legal@cyan.com
#include "HeadSpin.h" #include "HeadSpin.h"
#include "hsGeometry3.h" #include "hsGeometry3.h"
#include "hsCpuID.h"
class hsQuat; class hsQuat;
@ -104,7 +105,7 @@ struct hsMatrix44 {
const hsVector3* up); const hsVector3* up);
hsBool GetParity() const; hsBool GetParity() const;
float GetDeterminant() const; float GetDeterminant() const;
hsMatrix44* GetInverse(hsMatrix44* inverse) const; hsMatrix44* GetInverse(hsMatrix44* inverse) const;
hsMatrix44* GetTranspose(hsMatrix44* inverse) const; hsMatrix44* GetTranspose(hsMatrix44* inverse) const;
hsMatrix44* GetAdjoint(hsMatrix44* adjoint) const; hsMatrix44* GetAdjoint(hsMatrix44* adjoint) const;
@ -140,7 +141,7 @@ struct hsMatrix44 {
return rVal; return rVal;
} }
hsVector3 operator*(const hsVector3& p) const; hsVector3 operator*(const hsVector3& p) const;
hsMatrix44 operator*(const hsMatrix44& b) const; hsMatrix44 operator *(const hsMatrix44& other) const { return mat_mult.call(*this, other); }
hsPoint3* MapPoints(long count, hsPoint3 points[]) const; hsPoint3* MapPoints(long count, hsPoint3 points[]) const;
@ -152,6 +153,12 @@ struct hsMatrix44 {
void Read(hsStream *stream); void Read(hsStream *stream);
void Write(hsStream *stream); void Write(hsStream *stream);
// CPU-optimized functions
typedef hsMatrix44(*mat_mult_ptr)(const hsMatrix44&, const hsMatrix44&);
static hsMatrix44 mat_mult_fpu(const hsMatrix44&, const hsMatrix44&);
static hsMatrix44 mat_mult_sse3(const hsMatrix44&, const hsMatrix44&);
static hsFunctionDispatcher<mat_mult_ptr> mat_mult;
}; };
//////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////